ICDAR2024

The attached .pth file contains the weights of the pre-trained model referenced in the paper.
The configuration file can be found here.

Caution

The pretrained weights are not compatible with the latest commits in this repository.
Make sure to check out this release (5aaa5c5) exactly when using the provided model weights.

Training

Simply add the following line to the configuration file, and run train.py using the pretrained weights:

load_from = "/path/to/mutab_0dd0d49_pubtabnet_mutual_w300_im520.pth"

Inference

Prepare the ICDAR Task-B Test Data and run test.py using the pretrained weights as follows:

path=/path/to/data/icdar-task-b/final_eval
json=/path/to/data/icdar-task-b/final_eval.json
ckpt=/path/to/mutab_0dd0d49_pubtabnet_mutual_w300_im520.pth

python test.py --conf configs/pubtabnet.py --ckpt $ckpt --path $path --json $json

Annotation

The annotation format is fully compatible with MTL-TabNet and includes the following information: the image file name, HTML tags, and cell contents with their bounding boxes:

/path/to/pubtabnet/val/PMC4541863_007_00.png
<thead>,<tr>,<eb></eb>,<td></td>,</tr>,</thead>,<tbody>,<tr>,<td></td>,<td></td>,</tr>,<tr>,<td></td>,<td></td>,</tr>,</tbody>
0,0,0,0<;><UKN>
66,6,142,16<;>S	t	a	n	d	a	r	d	 	E	r	r	o	r
17,21,43,31<;>T	r	a	i	n
83,21,126,31<;>0	.	0	2	8	9	8	0
19,36,40,46<;>T	e	s	t
83,36,126,46<;>0	.	0	5	6	9	1	2